MDML_KD: Cross-Domain Knowledge Distillation Framework for Bearing Diagnosis

This paper introduces a novel multi-domain multi-level knowledge distillation (MDML-KD) framework designed to enhance bearing fault diagnosis in industrial environments, particularly under few-shot learning conditions. The proposed approach leverages multiple teacher models to transfer rich diagnostic knowledge to a lightweight student model, with intermediate assistant models acting as bridges to improve distillation efficiency. Extensive experiments on the CWRU, TBVD, and VDAB datasets demonstrate that the framework boosts student model accuracy by 1–7%, with the dual-teacher dual-assistant configuration achieving optimal performance under sufficient data. For secure edge deployment, the system is implemented on a Raspberry Pi 5, incorporating SHA-256-based key generation (derived from MAC addresses) and AES-CTR encryption to ensure robust data security. The results validate that the framework achieves high accuracy, computational efficiency, and secure communication, making it a viable solution for smart manufacturing and Industry 4.0 applications.